SecureVector

A official implementation of SecureVector Towards Privacy-Preserving, Real-Time and Lossless Feature Matching and involved baselines of template protection.

Usage

1. Download data for lfw/cfp/agedb from Gdrive or BaiduDrive.

2. Download IJB from BaiduDrive part1 and part2. Merge them by command cat data2a* > data2.tar.

3. Extract them in the root directory. You should have the following structure:

   Note: Features are extracted by MagFace. Replace the feat.list if you use another model.

    data/
    ├── agedb
    │ ├── agedb_feat.list
    │ └── pair.list
    ├── cfp
    │ ├── cfp_feat.list
    │ └── pair.list
    ├── ijb
    │ ├── ijbb_feat.list
    │ ├── ijbc_feat.list
    │ └── meta
    │     ├── ijbb_face_tid_mid.txt
    │     ├── ijbb_template_pair_label.txt
    │     ├── ijbc_face_tid_mid.txt
    │     └── ijbc_template_pair_label.txt
    └── lfw
    ├── lfw_feat.list
    └── pair.list

4. Run evaluations on the face task by:

    # [key] for method
        # 0. baseline
        # 1. SecureVector [1]
        # 2. ase [2]
        # 3. ironmask [3]
        # 4. sfm [4]

    export key=1 

    # LFW/CFP/AgeDB
    eval/eval1.sh $key  

    # IJB
    eval/evalibjx.sh $key

References

[1] Qiang Meng, el al, "Towards Privacy-Preserving, Real-Time and Lossless Feature Matching", arXiv 2022.

[2] Dusmanu, Mihai, et al. "Privacy-preserving image features via adversarial affine subspace embeddings." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

[3] Kim, Sunpill, et al. "Ironmask: Modular architecture for protecting deep face template." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

[4] Boddeti, Vishnu Naresh. "Secure face matching using fully homomorphic encryption." 2018 IEEE 9th International Conference on Biometrics Theory, Applications and Systems (BTAS). IEEE, 2018.